Depth Image-based Road Scene Segmentation And Target Tracking Technology

As the future development direction of automobiles,unmanned driving technology has a far-reaching impact on automobiles even the entire transportation industry.In recent years,with the development of environmental sensing sensors and artificial intelligence technology,unmanned driving technology has entered the stage of high-speed development and industrialization.Lidar is one of the most important environmental sensors used in driverless vehicles.The segmentation of 3D point cloud data collected by lidar and target tracking technology based on it can provide 3D environmental information for vehicle control system.The 3D point cloud data collected by multi-line lidar for unmanned vehicle has the characteristics of high noise and low resolution,and the computational complexity of three-dimensional space is relatively high.This paper uses the method of data resampling to project the point cloud onto the 2D plane to generate the depth image,and on this basis,analyses and designs the road segmentation technology based on depth histogram,the non-road partial segmentation technology using the correlation connection of adjacent pixels,and the target tracking technology using normalized depth histogram statistics.The main work of this paper is as follows:(1)The Cartesian coordinates of point cloud are converted into two-dimensional row and column index coordinates of image by data resampling method,and the distance between lidar and point is used as the value of pixel to form depth image.Aiming at the missing of some points,this paper uses bilinear interpolation to complete the image.Due to the technology designed in this paper is mainly used in urban road scenes,depth histogram statistics is used to carry out row-by-row statistics for depth images.The method of Random Sample Consensus is used to find the depth range of each corresponding road area,and the change threshold of depth value in each row is set to refine the depth range,so as to extract the road area.(2)For non-road areas,this paper uses 8-neighborhood method to scan sequentially,and uses the correlation of adjacent pixels as its segmentation basis to connect the related pixels to form connected areas.Finally,all connected regions are labeled one by one to generate the labeled image.(3)Based on the original depth image and the labeled image after segmentation,this paper takes camshaft algorithm as the frame model and counts the normalized histogram of each segmented block.Then initializing the tracking target in the first frame image and using its rectangular bounding box as the search window to find the segmentation block in the next frame image at the same position in a certain area around it.The depth histogram is used to calculate the chi-square similarity of two blocks in two frames,which can be used as a tracking basis to judge the correlation between the blocks and position relationship.So as to achieve the effect of target tracking.The experimental results on Kitti dataset verify the effectiveness of several techniques implemented in this paper.